A networked multi-unmanned autonomous robot cooperative tracking method and system

Abstract

The application relates to a networked multi-unmanned autonomous robot cooperative tracking method and system, and relates to the technical field of multi-unmanned autonomous robot cooperative tracking. A robot terminal collects real-time multi-source perception data of itself, processes the multi-source perception data, obtains a perception data set, and extracts features from the perception data set by using a trained feature extraction network. The features extracted by all robot terminals are input into a pre-constructed global feature fusion model, and a target feature vector is output. The target feature vector is input into a multi-constraint optimization model, the multi-constraint optimization model is solved by using a quantum particle swarm algorithm with a quantum rotation gate, and a cooperative tracking scheme is generated. The cooperative tracking scheme is verified by a digital twin engine in a physical field simulation, if all verifications meet preset safety and performance conditions, a plurality of instructions are generated based on the cooperative tracking scheme, and the instructions are sent to corresponding robot terminals. Otherwise, the target feature vector is regenerated, and real-time tracking of a dynamic target is realized.

Description

Technical Field

[0001] This invention relates to the field of collaborative tracking technology for multiple unmanned autonomous robots, specifically a networked collaborative tracking method and system for multiple unmanned autonomous robots. Background Technology

[0002] Multi-unmanned autonomous robot systems, such as drone swarms and unmanned vehicle fleets, have shown broad application prospects in many fields, including security patrols, logistics delivery, environmental monitoring, and disaster relief. Among these, collaborative tracking of dynamic targets is one of the core tasks of such systems. However, in practical applications, existing multi-unmanned autonomous robot collaborative tracking methods still face many challenges: First, system communication often suffers from high latency and limited bandwidth, resulting in untimely updates of decision-making information, which in turn affects the overall system's collaborative efficiency and tracking consistency.

[0003] Secondly, in complex dynamic environments, traditional tracking methods lack target tracking accuracy and struggle to effectively avoid path conflicts between robots.

[0004] Furthermore, for this type of high-dimensional, multi-constraint collaborative optimization problem, existing commonly used algorithms such as genetic algorithms and traditional particle swarm optimization often have slow convergence speeds, are prone to getting trapped in local optima, and have long computation times, making it difficult to meet the real-time response requirements of dynamic tasks.

[0005] Finally, most existing systems lack effective online learning and adaptive adjustment mechanisms, and cannot dynamically update their strategies based on changes in the environment, changes in target behavior, or abnormal states during task execution, resulting in poor robustness of the system in unstructured scenarios.

[0006] Therefore, there is an urgent need for a collaborative tracking method for multiple unmanned autonomous robots to improve the overall performance of multi-unmanned autonomous robot systems in real and complex scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a networked multi-unmanned autonomous robot cooperative tracking method and system, which enables real-time, accurate and low-cost tracking of dynamic targets, and improves the real-time performance, accuracy and robustness of multi-unmanned autonomous robot cooperative tracking.

[0008] To achieve the above objectives, the specific solution adopted by the present invention is as follows: a networked multi-unmanned autonomous robot collaborative tracking method, which employs a pre-built three-level networked architecture for networked multi-unmanned autonomous robot collaborative tracking. The three-level networked architecture includes a cloud platform, multiple edge computing nodes, and multiple robot terminals. The method includes the following steps: The robot terminal collects its own multi-source perception data in real time, preprocesses and weights it to obtain a perception dataset, and uses a trained feature extraction network to extract features from the perception dataset. Edge computing nodes acquire features extracted by all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs the target feature vector. Edge computing nodes acquire target feature vectors and input them into a multi-constraint optimization model. The quantum particle swarm optimization algorithm with quantum rotating gates is used to solve the multi-constraint optimization model and generate a cooperative tracking scheme. Edge computing nodes use a digital twin engine to perform physical field simulation verification of the collaborative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple instructions are generated based on the collaborative tracking scheme and sent to the corresponding robot terminals; otherwise, the target feature vector is regenerated. The edge computing node calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters based on the deviation, and sends the robot optimization parameters to the robot terminal.

[0009] As an optimization of the aforementioned networked multi-unmanned autonomous robot collaborative tracking method, the method for real-time collection of multi-source perception data by the robot terminal and its preprocessing and weighting is as follows: The robot terminal collects key data and environmental auxiliary data every 100ms; After removing outliers from key data and environmental auxiliary data using the 3σ criterion, the remaining key data and environmental auxiliary data are mapped to the [0,1] interval for normalization. An attention mechanism is introduced to weight the normalized key data and environmental auxiliary data. The weight of the key data is 0.6 and the weight of the environmental auxiliary data is 0.2. The weights are normalized by the Softmax function to obtain the perception dataset.

[0010] As another optimization scheme for the above-mentioned networked multi-unmanned autonomous robot collaborative tracking method: when the deviation is greater than or equal to 0.25, the weight of key data is adjusted to 0.7, the weight of environmental auxiliary data is adjusted to 0.3, and the weight is updated by gradient descent method with a learning rate of 0.003.

[0011] As another optimization scheme for the above-mentioned networked multi-unmanned autonomous robot cooperative tracking method, the trained feature extraction network is obtained by training a ResNet18 sub-model.

[0012] As an alternative optimization of the aforementioned networked multi-unmanned autonomous robot collaborative tracking method, the global feature fusion model construction process involves using a federated averaging algorithm to aggregate the parameters of the feature extraction network in each robot terminal, and generating a global feature fusion model at the edge computing nodes.

[0013] As another optimization scheme of the above-mentioned networked multi-unmanned autonomous robot cooperative tracking method: the target feature vector is 32-dimensional, including the target's three-dimensional coordinates (x, y, z), motion velocity (vx, vy, vz), acceleration (ax, ay, az) and feature confidence.

[0014] As an alternative optimization of the aforementioned networked multi-unmanned autonomous robot cooperative tracking method, the method for generating the cooperative tracking scheme is as follows: The target feature vector is input into a multi-constraint optimization model, and the constraints of the multi-constraint optimization model are: robot speed ≤ 5 m / s, acceleration ≤ 2 m / s². 2 Communication time slot ≤ 10ms; Initialize the particle swarm, with a particle size of 80, and each particle corresponds to a cooperative tracking scheme; Design a fitness function; Introducing a quantum rotation gate to update particle positions; The algorithm iterates until the maximum number of iterations (80) is met. When the fitness function fluctuation is less than 0.01 for 5 consecutive iterations, the algorithm converges. During the iteration process, the historical optimal solution and the global optimal solution for each particle are recorded. Every 10 iterations, the current optimal solution is output. After convergence, the optimal solution is output as the cooperative tracking solution.

[0015] As an alternative optimization scheme for the aforementioned networked multi-unmanned autonomous robot cooperative tracking method, the deviation is: , in, Let Euclidean distance be the actual position of the robot and its predicted position. The percentage difference between the actual speed and the predicted speed. It represents the percentage difference between actual energy consumption and predicted energy consumption. , , These are the weighting coefficients.

[0016] As an alternative optimization of the aforementioned networked multi-unmanned autonomous robot cooperative tracking method, physical field simulation verification includes: Kinematic verification: A virtual environment with a 1:1 scale to the actual scene is built in Unity3D to simulate the robot's tracking trajectory. If the path deviation is less than 0.3mm and the speed fluctuation is less than 5%, the condition is met. Dynamic verification was performed by ADAMS simulation of robot joint torque. If the torque fluctuation is <10 N·m and the acceleration impact is <2 m / s2, then the conditions are met. Electromagnetic compatibility (EMC) verification is performed using a simulated communication module. If the electromagnetic interference intensity is less than 2V / m, the condition is met.

[0017] A networked multi-unmanned autonomous robot collaborative tracking system includes a cloud platform, multiple robot terminals, and multiple edge computing nodes; The robot terminal includes a feature extraction module and multiple sensors. The sensors collect their own multi-source perception data in real time, and perform preprocessing and weighted processing to obtain a perception dataset. The trained feature extraction network is then used to extract features from the perception dataset. Edge computing nodes include a feature fusion module, a collaborative decision-making module, a dynamic adjustment module, and a simulation verification module; The feature fusion module is used to acquire the features extracted by all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs the target feature vector. The collaborative decision-making module is used to obtain the target feature vector and input it into the multi-constraint optimization model. The quantum particle swarm optimization algorithm with quantum rotating gate is used to solve the multi-constraint optimization model and generate a collaborative tracking scheme. The simulation verification module uses a digital twin engine to perform physical field simulation verification of the cooperative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple instructions are generated based on the cooperative tracking scheme and sent to the corresponding robot terminal; otherwise, the target feature vector is regenerated. The dynamic adjustment module calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters based on the deviation, and sends the robot optimization parameters to the robot terminal. The cloud platform includes a feedback and iteration module.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a networked multi-unmanned autonomous robot cooperative tracking method and system. It adopts a three-level networked architecture consisting of a cloud platform, multiple edge computing nodes, and multiple robot terminals, and introduces the quantum particle swarm optimization algorithm with quantum rotating gate, so that the cooperative tracking response time is no more than 200ms, which is 60% shorter than the traditional method, meets the requirements of dynamic target tracking, and improves the real-time performance of the method.

[0019] The features extracted from each robot terminal are fused to achieve a target position error of ≤0.3mm, a recognition accuracy of ≥95%, and a target loss rate of ≤1%, which is 40% higher than traditional methods and improves tracking accuracy.

[0020] A multi-constraint optimization model is used to generate a collaborative tracking scheme, which reduces total energy consumption by 25%, ensures path overlap rate of ≤5%, avoids robot collisions and resource waste, and optimizes robot energy consumption and tracking path.

[0021] By employing deviation and iterative updates, the tracking recovery time is ≤1s and the recovery rate is ≥98% in scenarios involving sudden changes in the target trajectory and robot failures, thus enhancing dynamic adaptability. Attached Figure Description

[0022] Figure 1 This is a flowchart of the collaborative tracking method of the present invention.

[0023] Figure 2 This is a schematic diagram of a three-tier network architecture.

[0024] Figure 3 This is a diagram of the quantum particle swarm optimization algorithm. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0026] A networked multi-unmanned autonomous robot collaborative tracking method is proposed, employing a pre-built three-tiered network architecture. This architecture includes a cloud platform, multiple edge computing nodes, and multiple robot terminals. Specifically, the cloud platform is deployed in a remote data center, trained using a long-term trajectory prediction model, and possesses elastic scaling capabilities, supporting elastic expansion up to 32 cores and 64GB of memory. It is responsible for global trajectory prediction based on long-term historical data, historical data storage for one year, and management and updating of the networked learning global model. The edge computing nodes utilize 5G edge gateways, deployed around the tracking area at intervals of 1 km. 2One robot is deployed at a regional density, responsible for real-time data fusion processing, short-term tracking scheduling within 5 minutes, and high-frequency communication with the robot terminal. Data throughput is ≥100Mbps, and short-term scheduling for tracking scheme adjustments within 5 minutes is performed, with communication latency ≤50ms. The robot terminal is equipped with multi-source sensors and a 5G communication module, responsible for collecting multi-source perception data (target and environmental data), performing tracking quality checks, and providing feedback on the robot's status information. In this embodiment, the robot terminal uses an AGV-T10 unmanned robot, equipped with a 5G module (SDX55), a 16-line LiDAR, a 2-megapixel vision sensor, and a 24GHz millimeter-wave radar. It uploads multi-source perception data every 100ms, and the response time for executing tracking commands is ≤30ms. The communication protocol adopts a 5G NR + edge computing fusion protocol. The cloud platform and edge computing nodes transmit via 10Gbps single-mode fiber optic cable, while the edge computing nodes and the robot communicate via millimeter-wave communication with a coverage radius of 1km. The packet loss rate is ≤0.1%, meeting the communication requirements for real-time tracking.

[0027] The method includes the following steps: Multi-source sensory data fusion is performed based on federated learning and attention mechanisms. The robot terminal collects its own multi-source sensory data in real time, and performs outlier removal, normalization preprocessing, and weighting. An attention mechanism is introduced to differentially weight key data such as target position and velocity with environmental auxiliary data, and then normalizes the data using the Softmax function to obtain the sensory dataset. A pre-trained feature extraction network is then used to extract features from the sensory dataset. Specifically: The robot terminal collects key data and environmental auxiliary data every 100ms.

[0028] After removing outliers caused by lidar occlusion in key data and environmental auxiliary data using the 3σ criterion, the remaining key data and environmental auxiliary data are mapped to the [0,1] interval for normalization to avoid the influence of dimensions.

[0029] An attention mechanism is introduced to weight the normalized key data and environmental auxiliary data. The key data (target position, target velocity) has a weight of 0.6, and the environmental auxiliary data (obstacle distance, light intensity) has a weight of 0.2. The weights are normalized by the Softmax function to obtain the perception dataset, ensuring that the core tracking information is focused.

[0030] A pre-trained feature extraction network is used to extract features from the perception dataset. Specifically, the feature extraction network for each robot terminal is trained based on a ResNet18 sub-model. Edge computing nodes acquire features extracted from all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs a target feature vector. The parameters of the feature extraction network in each robot terminal are aggregated using a federated averaging algorithm, and a global feature fusion model is generated at the edge computing nodes. A maximum mean difference (MMD) constraint is applied to ensure the consistency of feature distribution across robot terminals, with an MMD value ≤ 0.05, aligning feature distributions and avoiding fusion bias caused by heterogeneous perceptual datasets. The federated averaging algorithm is as follows: , Where, n i The amount of data in the robot's terminal perception dataset.

[0031] The target feature vector is 32-dimensional, including the target's three-dimensional coordinates (x, y, z), motion velocity (vx, vy, vz), acceleration (ax, ay, az), and feature confidence. The accuracy of the target's three-dimensional coordinates is 0.1m, and the motion velocity ranges from 0.1 to 10m / s. This provides accurate input for subsequent optimization.

[0032] Edge computing nodes acquire target feature vectors and input them into a multi-constraint optimization model. A quantum particle swarm optimization algorithm incorporating quantum rotation gates is then used to solve the multi-constraint optimization model, generating a cooperative tracking scheme. Specifically: The target feature vector is input into a multi-constraint optimization model. The multi-constraint optimization model is constructed with the objectives of "highest tracking accuracy, lowest total energy consumption, and minimum path overlap rate". The constraints of the multi-constraint optimization model are: robot speed ≤ 5m / s and acceleration ≤ 2m / s². 2 The constraint is that the communication time slot is ≤10ms.

[0033] Initialize a particle swarm with 80 particles, each particle corresponding to a cooperative tracking scheme. Ion positions are encoded using real numbers, and each particle is encoded as a robot ID (1-N), tracking path node, motion speed, and communication time slot. Path nodes use latitude and longitude coordinates with an accuracy of 0.0001°, motion speeds range from 0.1 to 5 m / s, and communication time slots range from 1 to 10 ms.

[0034] Design a fitness function, the fitness function is: ; Where P is the reciprocal of the target position error. , E represents the location error; E represents the total energy consumption; and O represents the path repetition rate (0-1). , and These are the weighting coefficients, and + + =1, initial value =0.5、 =0.3 and =0.2.

[0035] Introducing a quantum rotation gate to update particle position and rotation angle Set the range to [-π / 4, π / 4], and update the formula: ;in, Let i be the position of particle i in the t-th iteration. Let be the phase angle of particle i. The quantum superposition property is used to enhance the global search capability and avoid getting trapped in local optima. For velocity updates, combined with robot motion constraints, the velocity adjustment amount is ≤0.5m / s to avoid overshoot.

[0036] The algorithm iterates until the maximum number of iterations (80) is met. When the fitness function fluctuation is less than 0.01 for 5 consecutive iterations, the algorithm converges. During the iteration process, the historical optimal solution and the global optimal solution for each particle are recorded. The current optimal solution is output every 10 iterations. After convergence, the optimal solution is output as the cooperative tracking solution. The solution time is ≤200ms, which meets the real-time scheduling requirements.

[0037] Edge computing nodes use a digital twin engine to perform physical field simulation verification of the cooperative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple commands are generated based on the cooperative tracking scheme and sent to the corresponding robot terminals. For example, the edge computing node sends the tracking command "Robot 1 moves at a speed of 2 m / s along the particle size point (30.1234°, 120.5678°)" to the robot terminal, with a command transmission delay of no more than 30 ms. Otherwise, the target feature vector is regenerated, and the regenerated target feature vector is re-input into the multi-constraint optimization model to regenerate the cooperative tracking scheme and perform verification. Specifically, the physical field simulation verification includes: Kinematic verification: A virtual environment with a 1:1 scale to the actual scene is built in Unity3D to simulate the robot's tracking trajectory. If the path deviation is less than 0.3mm and the speed fluctuation is less than 5%, the condition is met. Dynamic verification was performed using ADAMS simulation of robot joint torques. If torque fluctuations were <10 N·m and acceleration impacts <2 m / s², the results were satisfactory. 2 If so, then the condition is met; Electromagnetic compatibility (EMC) verification is performed using a simulated communication module. If the electromagnetic interference intensity is less than 2V / m, the condition is met.

[0038] The edge computing node calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters based on the deviation, and sends the robot optimization parameters to the robot terminal.

[0039] The deviation is quantified as follows: , in, Let Euclidean distance be the actual position of the robot and its predicted position. The percentage difference between the actual speed and the predicted speed. It represents the percentage difference between actual energy consumption and predicted energy consumption. , , These are the weighting coefficients.

[0040] when At that time, maintain the current parameters.

[0041] when Fine-tune the robot speed (±0.2m / s) and the federated learning weight update rate (0.01→0.02).

[0042] when At that time, the acceleration was adjusted significantly (±0.5m / s²). 2 Increase the weight update rate to 0.05 to prioritize tracking accuracy.

[0043] when At that time, the weight of key data was adjusted to 0.7, the weight of environmental auxiliary data was adjusted to 0.3, and the weights were updated using gradient descent with a learning rate of 0.003.

[0044] When the deviation is high, increase the tracking accuracy weight in the fitness function. Reduce the energy consumption weight to 0.6. The value is reduced to 0.1, effectively ensuring that the target is not lost.

[0045] The communication strategy is optimized based on the deviation degree. The TDMA protocol is used to reallocate time slots, allocating 5ms short time slots to robots with high deviation degrees, thereby increasing the data upload frequency and ensuring real-time adjustment.

[0046] The aforementioned method constructs a closed-loop system integrating hierarchical computing power allocation, dual-mode communication redundancy, privacy-preserving perception fusion, quantum-inspired global optimization, digital twin forward-looking verification, and multi-dimensional deviation feedback iteration. First, a three-level "cloud-edge-device" network architecture is built, achieving low-latency data interaction and computing power scheduling among multiple robots through hierarchical computing power allocation and dual-mode communication redundancy. Then, multi-source perception data is collected through multi-source heterogeneous sensors on the robot terminal. Based on federated learning combined with a Softmax weight normalization attention mechanism, adaptive weighted fusion of multi-source heterogeneous perception data is achieved while protecting data privacy. Next, the collaborative tracking problem is modeled as a multi-objective optimization problem, and an improved particle swarm optimization algorithm incorporating a quantum rotating gate update mechanism is used for efficient global solution, balancing tracking accuracy, system energy consumption, and path conflict. Finally, kinematic, dynamic, and electromagnetic compatibility multi-physics simulation verification is performed using a digital twin engine. Based on the multi-dimensional tracking deviation fusion of position, velocity, and energy consumption, the robot's motion parameters, model weights, and communication strategies are dynamically adjusted, and system performance is continuously iteratively optimized through feedback data.

[0047] A networked multi-unmanned autonomous robot collaborative tracking system includes a cloud platform, multiple robot terminals, and multiple edge computing nodes; it deploys 5G communication modules and edge gateways to achieve data interaction with a latency of no more than 50ms and computing power allocation of no less than 10TOPS for edge nodes.

[0048] The robot terminal includes a feature extraction module and multiple sensors, which are integrated vision, 16-line LiDAR, and 24GHz millimeter-wave radar sensors with a resolution of 1920×1080. The sensors collect their own multi-source perception data in real time, and perform preprocessing and weighted processing to obtain a perception dataset. The trained feature extraction network is then used to extract features from the perception dataset. Edge computing nodes include a feature fusion module, a collaborative decision-making module, a dynamic adjustment module, and a simulation verification module; The feature fusion module is used to acquire the features extracted by all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs the target feature vector. The collaborative decision-making module is equipped with a quantum particle swarm optimization algorithm solver, which supports dynamic adjustment of particle swarm size, quantum rotation gate update and iterative determination. The collaborative decision-making module is mainly used to obtain the target feature vector and input it into the multi-constraint optimization model. The quantum particle swarm optimization algorithm with quantum rotation gate is used to solve the multi-constraint optimization model and generate a collaborative tracking scheme.

[0049] The simulation verification module uses a digital twin engine to perform physical field simulation verification of the collaborative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple instructions are generated based on the collaborative tracking scheme and sent to the corresponding robot terminals; otherwise, the target feature vector is regenerated.

[0050] The dynamic adjustment module calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters according to the deviation, and sends the robot optimization parameters to the robot terminal; it also supports sensor fault detection and switching.

[0051] The cloud platform includes a feedback iteration module that receives real-time feedback data from the robot terminal with a sampling frequency of ≥10Hz. The data includes: robot position, speed, energy consumption, and sensor status. Iterative updates and optimizations are performed on model parameters, communication strategies, and federated learning models.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A networked multi-unmanned autonomous robot cooperative tracking method, characterized in that: The method employs a pre-established three-tiered network architecture for collaborative tracking of multiple unmanned autonomous robots. This three-tiered network architecture includes a cloud platform, multiple edge computing nodes, and multiple robot terminals. The method includes the following steps: The robot terminal collects its own multi-source perception data in real time, preprocesses and weights it to obtain a perception dataset, and uses a trained feature extraction network to extract features from the perception dataset. Edge computing nodes acquire features extracted by all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs the target feature vector. Edge computing nodes acquire target feature vectors and input them into a multi-constraint optimization model. The quantum particle swarm optimization algorithm with quantum rotating gates is used to solve the multi-constraint optimization model and generate a cooperative tracking scheme. Edge computing nodes use a digital twin engine to perform physical field simulation verification of the collaborative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple instructions are generated based on the collaborative tracking scheme and sent to the corresponding robot terminals; otherwise, the target feature vector is regenerated. The edge computing node calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters based on the deviation, and sends the robot optimization parameters to the robot terminal.

2. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: The method for the robot terminal to collect its own multi-source perception data in real time and to preprocess and weight it is as follows: The robot terminal collects key data and environmental auxiliary data every 100ms; After removing outliers from key data and environmental auxiliary data using the 3σ criterion, the remaining key data and environmental auxiliary data are mapped to the [0,1] interval for normalization. An attention mechanism is introduced to weight the normalized key data and environmental auxiliary data. The weight of the key data is 0.6 and the weight of the environmental auxiliary data is 0.

2. The weights are normalized by the Softmax function to obtain the perception dataset.

3. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 2, characterized in that: When the bias is greater than or equal to 0.25, the weight of the key data is adjusted to 0.7, the weight of the environmental auxiliary data is adjusted to 0.3, and the weights are updated using the gradient descent method with a learning rate of 0.

003.

4. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: The trained feature extraction network is obtained by training a ResNet18 sub-model.

5. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: The global feature fusion model is constructed by using a federated averaging algorithm to aggregate the parameters of the feature extraction network in each robot terminal and generating a global feature fusion model at the edge computing nodes.

6. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: The target feature vector is 32-dimensional, including the target's three-dimensional coordinates (x, y, z), motion velocity (vx, vy, vz), acceleration (ax, ay, az), and feature confidence.

7. The networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: The method for generating a collaborative tracking scheme is as follows: The target feature vector is input into a multi-constraint optimization model, and the constraints of the multi-constraint optimization model are: robot speed ≤ 5 m / s, acceleration ≤ 2 m / s². 2 Communication time slot ≤ 10ms; Initialize the particle swarm, with a particle size of 80, and each particle corresponds to a cooperative tracking scheme; Design a fitness function; Introducing a quantum rotation gate to update particle positions; The algorithm iterates until the maximum number of iterations (80) is met. When the fitness function fluctuation is less than 0.01 for 5 consecutive iterations, the algorithm converges. During the iteration process, the historical optimal solution and the global optimal solution for each particle are recorded. Every 10 iterations, the current optimal solution is output. After convergence, the optimal solution is output as the cooperative tracking solution.

8. A networked multi-unmanned autonomous robot cooperative tracking method as described in claim 7, characterized in that, The deviation is: ， in, Let Euclidean distance be the actual position of the robot and its predicted position. The percentage difference between the actual speed and the predicted speed. It represents the percentage difference between actual energy consumption and predicted energy consumption. , , These are the weighting coefficients.

9. A networked multi-unmanned autonomous robot cooperative tracking method as described in claim 1, characterized in that: Physics field simulation verification includes: Kinematic verification: A virtual environment with a 1:1 scale to the actual scene is built in Unity3D to simulate the robot's tracking trajectory. If the path deviation is less than 0.3mm and the speed fluctuation is less than 5%, the condition is met. Dynamic verification was performed using ADAMS simulation of robot joint torques. If torque fluctuations were <10 N·m and acceleration impacts <2 m / s², the results were satisfactory. 2 If so, then the condition is met; Electromagnetic compatibility (EMC) verification is performed using a simulated communication module. If the electromagnetic interference intensity is less than 2V / m, the condition is met.

10. A networked multi-unmanned autonomous robot collaborative tracking system, characterized in that: This includes a cloud platform, multiple robot terminals, and multiple edge computing nodes; The robot terminal includes a feature extraction module and multiple sensors. The sensors collect their own multi-source perception data in real time, and perform preprocessing and weighted processing to obtain a perception dataset. The trained feature extraction network is then used to extract features from the perception dataset. Edge computing nodes include a feature fusion module, a collaborative decision-making module, a dynamic adjustment module, and a simulation verification module; The feature fusion module is used to acquire the features extracted by all robot terminals and input them into a pre-built global feature fusion model. The global feature fusion model outputs the target feature vector. The collaborative decision-making module is used to obtain the target feature vector and input it into the multi-constraint optimization model. The quantum particle swarm optimization algorithm with quantum rotating gate is used to solve the multi-constraint optimization model and generate a collaborative tracking scheme. The simulation verification module uses a digital twin engine to perform physical field simulation verification of the cooperative tracking scheme. If all verifications meet the preset safety and performance conditions, multiple instructions are generated based on the cooperative tracking scheme and sent to the corresponding robot terminal; otherwise, the target feature vector is regenerated. The dynamic adjustment module calculates the deviation based on the robot's actual tracking results and the generated collaborative tracking scheme, and dynamically generates its own optimization parameters and robot optimization parameters based on the deviation, and sends the robot optimization parameters to the robot terminal. The cloud platform includes a feedback and iteration module.

Images